A couple of years ago I spent real time collecting "magic phrases." Tell the model it's an expert with twenty years of experience. Tell it to think step by step. Promise it a tip if it does a good job. Format the prompt just so, with the right delimiters and the right order of instructions, because if you got the wording wrong the answer would be noticeably worse. That was prompt engineering, and for a while it was a legitimate skill worth paying someone to teach you. I don't think it is anymore, and I want to explain why, because a lot of the content still being sold to small business owners is built around a problem that's mostly gone away.
The reason those tricks worked is that earlier models were genuinely brittle. They'd latch onto surface patterns in your wording, get thrown off by ambiguous phrasing, lose track of the actual question over a long answer, or need to be coaxed into reasoning carefully instead of blurting out the first plausible-sounding response. Coaxing a weaker model into its best behavior with the right incantation was a real, learnable, and valuable skill. It's the reason "prompt engineer" was briefly a job title.
Today's stronger models just don't need that kind of coaxing nearly as much. Tell a current model plainly what you want, in normal sentences, the way you'd explain it to a competent new hire, and it usually does a good job. The gap between a cleverly engineered prompt and a plainly worded one has shrunk enormously, and for most everyday business tasks it's closed almost entirely. Clear instructions still matter — vague or contradictory requests still get you vague or contradictory answers — but that's just clear communication, not a special incantation. The wordsmithing arms race is largely over.
I had a client last year who was convinced her marketing copy problem was a prompting problem. She'd collected a folder of prompt templates from a paid course, tweaking phrases, adding roleplay framing, trying different "personas" for the model to adopt. None of it changed much, because the actual problem wasn't how she was asking — it was that the model had never seen her pricing sheet, her actual customer complaints, or the three competitor sites she was trying to differentiate from. Once we fed it that material directly, the same plain-English request that had been failing for weeks started producing usable drafts on the first try.
That's the shift in one sentence: what determines answer quality now is mostly what information you give the model to work with, not how you phrase the ask. People have started calling this context engineering, which is a slightly grander name for a fairly simple idea — the model can only reason well about what's actually in front of it. Your invoicing history, your actual customer emails, your specific product catalog, the document that has the real numbers in it. Assembling and handing over the right material, in a form the model can actually use, does far more for output quality than any clever phrasing ever did.
The second thing that matters more now is picking the right tool or mode for the job, because "AI" stopped being one thing a while ago. A quick chat window is fine for drafting an email. It's the wrong tool for something that needs to search the live web, or dig through fifty pages of a contract, or actually take an action in your calendar or your accounting software. Most of the AI products businesses use today ship with multiple modes — a fast conversational mode, a slower research mode that goes and gathers sources, an agent mode that can use tools and take multi-step action. Getting a good result increasingly depends on recognizing which mode a task actually calls for, not on how you word the request inside the wrong mode.
The third thing is structure — how you break a real piece of work into steps instead of expecting one message to produce a finished result. Complex work goes better when you split it: gather the facts first, then draft, then check the draft against the facts, then revise. That's true whether a person or a model is doing the work, and it was true before AI existed too. What's changed is that current models are good enough to be trusted to execute each of those steps well, so the leverage has moved to whoever designs the sequence of steps and checks the work along the way, rather than whoever finds the cleverest single sentence to kick things off.
Put those three together and you get a pretty different picture of what "being good at AI" means for a business owner in practice. It's less about learning a vocabulary of magic words and more about three ordinary skills: knowing what information actually matters and making sure the model has it, knowing which tool or mode fits the job at hand, and knowing how to break a real piece of work into a sequence that a model can execute reliably. None of those are exotic. They're closer to the skills of a decent manager delegating work than the skills of a hacker finding an exploit.
If you're a small business owner who bought into the idea that you needed to master prompt engineering to get value from AI, I'd stop worrying about that specific thing. Don't spend more time hunting for the perfect phrasing. Spend it organizing the information in your business so it's actually accessible — your pricing, your policies, your past customer interactions, whatever's specific to how you operate — because that's what a model actually needs to give you a good answer instead of a generic one. Spend it figuring out which tool actually fits a given task instead of using the one tab you always have open for everything. And when a task is genuinely multi-step, take the ten minutes to break it into stages rather than expecting one message to do it all.
None of this means plain instructions don't matter at all — being specific about what you want, what "good" looks like, and what constraints apply still makes a real difference. It's just that this is ordinary clear communication, the same thing that makes you a better manager or a better writer of instructions to a human employee, not a hidden trick you need a course to unlock. The skill that actually separates people getting real value from AI from people who are frustrated by it isn't wordsmithing anymore. It's judgment about information, tools, and process — which, not coincidentally, is exactly the kind of judgment a business owner already has about their own business, if someone helps them apply it.
I write about this kind of thing at 013labs.com, mostly because I keep meeting business owners who are stuck on the wrong problem — still hunting for magic words when the real leverage has moved somewhere else entirely.